Bibliographic study showed improving statistical methodology of network meta-analyses published between 1999 and 2015

To assess the characteristics and core statistical methodology specific to network meta-analyses (NMAs) in clinical research articles.

We searched MEDLINE, EMBASE, and the Cochrane Database of Systematic Reviews from inception until April 14, 2015, for NMAs of randomized controlled trials including at least four different interventions. Two reviewers independently screened potential studies, whereas data abstraction was performed by a single reviewer and verified by a second.

A total of 456 NMAs, which included a median (interquartile range) of 21 (13-40) studies and 7 (5-9) treatment nodes, were assessed. A total of 125 NMAs (27%) were star networks; this proportion declined from 100% in 2005 to 19% in 2015 (P = 0.01 by test of trend). An increasing number of NMAs discussed transitivity or inconsistency (0% in 2005, 86% in 2015, P < 0.01) and 150 (45%) used appropriate methods to test for inconsistency (14% in 2006, 74% in 2015, P < 0.01). Heterogeneity was explored in 256 NMAs (56%), with no change over time (P = 0.10). All pairwise effects were reported in 234 NMAs (51%), with some increase over time (P = 0.02). The hierarchy of treatments was presented in 195 NMAs (43%), the probability of being best was most commonly reported (137 NMAs, 70%), but use of surface under the cumulative ranking curves increased steeply (0% in 2005, 33% in 2015, P < 0.01).

Many NMAs published in the medical literature have significant limitations in both the conduct and reporting of the statistical analysis and numerical results. The situation has, however, improved in recent years, in particular with respect to the evaluation of the underlying assumptions, but considerable room for further improvements remains.